Speaker
Description
The organization of shallow convection in the trades influences precipitation development, cold pool formation, clouds’ radiative effects, and, thus, climate sensitivity. Deep learning techniques, especially in computer vision and self-supervision, are suitable tools to understand cloud organization purely from a machine’s perspective. Yet, the physical interpretation of the network’s classes remains challenging.
We use a self-supervised deep learning neural network to assess the organization and the associated physical properties of shallow convection during the EUREC4A field study. Based on geostationary GOES-East images of cloud optical depth, we analyze images from both Barbados Cloud Observatory, and from a random selection of a larger domain across the tropical Atlantic. The network classifies each image into one of four classes of organization. Based on the neural network outputs, that is higher-order image features and corresponding labels of the image, we analyze how these relate to the four categories previously identified from human labeling by the EUREC4A community (Sugar, Gravel, Flower, Fish).
